This is a submission for the Notion MCP Challenge

What I Built

I maintain a couple of open source projects on my own, and the worst part has never been writing code — it's the stuff around it. Going through new issues, deciding if something's a bug or a feature request, digging into the codebase to figure out what's actually wrong, and then sitting down to fix it. Half the issues end up sitting untouched for weeks because I just don't get to them.

So I built Ghost Maintainer. It turns Notion into an operations center for a GitHub repo, and handles most of the grunt work with AI. Here's what happens when someone opens an issue:

1. It shows up in a Triage Queue in Notion
2. Gemini reads the issue and classifies it — bug or feature, priority level, confidence score
3. Bugs go to the Maintenance Backlog, features go to the Feature Backlog
4. If auto-fix is on, it reads through the codebase, figures out a fix, and opens a PR

All I have to do is review the PR and merge.

If you've turned off auto-fix (or want to handle features), there's a CLI:

ghost_maintainer fix 3
ghost_maintainer implement 9

Same pipeline — reads the code, writes a fix, opens a PR.

Setting it up is one command from inside your repo:

curl -sL https://raw.githubusercontent.com/sbis04/ghost_maintainer/main/install.sh \
  -o install.sh && \
  bash install.sh && \
  rm install.sh

You plug in your Notion, GitHub, and Gemini tokens, and it handles the rest — creates the databases, sets up GitHub secrets, pushes the workflows.

There's also a Dart MCP server that pairs with Notion MCP (@notionhq/notion-mcp-server). Hook both up to Gemini CLI or Claude, and you can interact with your backlog in plain English — triage stuff, investigate bugs, ship fixes. 5 tools, 2 resources, 2 prompts.

The Notion workspace

• Triage Queue — everything starts here, AI sorts it out
• Maintenance Backlog — bugs, with AI summaries and PR links
• Feature Backlog — feature requests, ready when you are
• Archive — done items land here with timestamps
• Project Vision Statement — tells the AI what matters to your project

Video Demo

Link to Code on GitHub

Three main pieces:

• ghost_maintainer_mcp/ — Dart MCP server. 5 tools, 2 resources, 2 prompts.
• notion_setup/ — CLI tool + the GitHub Action scripts.
• .github/workflows/ — 4 workflows covering ingestion, triage, implementation, and archival.

The CLI installs globally via dart pub global activate straight from the repo — no need to publish to pub.dev.

How I Used Notion MCP

The whole system runs on two MCP servers working together.

Notion MCP (@notionhq/notion-mcp-server) is the foundation. It's what gives the AI access to the Notion workspace — searching databases, reading pages, creating content, appending blocks. Without it, the AI would be blind to what's going on.

Ghost Maintainer MCP sits on top of that. It's a custom Dart server that adds the maintenance-specific stuff — ghost_triage_issue, ghost_investigate_issue, ghost_deploy_fix. The prompts are written to use both servers together. The triage prompt, for instance, tells the AI to search for duplicates through Notion MCP before classifying, and to append its analysis notes to the Notion page afterward.

Ghost Maintainer also exposes a couple of resources that pull from Notion — ghost://vision grabs the Project Vision Statement, and ghost://backlog/summary aggregates stats across all the databases. These save the AI from having to navigate the workspace manually every time.

Everything lives in Notion. The Triage Queue, backlogs, and archive are all databases with properties like Stage, Priority, Labels, AI Summary, and PR URL. The Vision Statement is a regular Notion page — update it and the AI adjusts. If you write "security issues are always P0" in there, that's exactly what happens.

PRs link back to the Notion page. When they get merged, the item auto-archives with a date.

The GitHub Actions pipeline handles the automated side (new issue triggers triage, etc.), but the MCP layer is what makes it interactive. Connect both servers to Gemini CLI and you get conversations like:

You: "What's in the backlog?"
AI: [uses ghost://backlog/summary] "3 bugs, 2 features..."

You: "Any duplicates for the dark mode request?"
AI: [uses Notion MCP search] "Found a similar one from last week."

You: "Triage the login crash"
AI: [uses ghost_triage_issue] "Bug, P1-High."

You: "Fix it"
AI: [uses ghost_investigate_issue + ghost_deploy_fix] "PR #15 created."

The thing I like about this is that Notion ends up being the shared workspace for both me and the AI. I open Notion and see what's going on. The AI checks the same workspace through Notion MCP. Same board, same data, no surprises.